Variant interpretation moves into the hospital server room: a 9B local model writes ACMG reports and recovers 93% of causal variants in 327 singleton cases

The reasoning step of variant interpretation moves to a local small model, which assigns ACMG/AMP criteria, reconciles inheritance with zygosity and parental genotypes, and writes a complete case-level report. Most existing systems leave classification to a deterministic engine and confine the language model to retrieving, extracting, compacting, and summarizing evidence, or handle one criterion or one variant at a time. Across 16 solved trios, IVA reported every previously identified pathogenic, likely pathogenic, or uncertain-significance variant of interest, with 100% sensitivity, 39.5% precision, and 98.2% accuracy over all candidate variants.

The pipeline brackets model reasoning with deterministic code that retrieves evidence, computes segregation and phase from allelic balance, recomputes point totals after every layer, validates criterion eligibility, and checks citations. Small models are unreliable at arithmetic, verbatim copying, and strict formatting, so those tasks were removed from the model wherever possible. In 327 singleton visual-impairment cases, IVA surfaced 93% of expert-reported variants, classified 82% as pathogenic or likely pathogenic and 11% as variants of uncertain significance of interest, with 43% precision and 99% accuracy over all candidate variants.

Model selection came from a 150-question clinical-genomics benchmark run under all 24 answer orderings per question, and settled on Qwen3.5-9B. The 1.6-trillion-parameter cloud model DeepSeek-V4-Pro scored higher on every metric, with mean accuracy 0.887 against 0.812. Thirteen local models from 0.8B to 9B parameters were evaluated on one A100, and Qwen3.5-9B reached the highest mean accuracy 0.812, strict accuracy 0.633, choice consistency 0.815, and inverted Brier score 0.875.

The report addresses researchers and engineers in clinical genetics and genomic diagnostics, covering case-level interpretation for singletons and trios, with an annotated variant table plus free-text phenotype as input and all model inference running locally on one A100. In trios the model applied de novo, dominant, recessive, and X-linked inheritance layers, and when phase could not be determined it reported compound heterozygosity as a hypothesis requiring confirmation rather than an established finding. The authors propose external validation, participation in quality assessment schemes such as EMQN, comparison with analysts, and distillation or quantization for less powerful hardware as next steps.

The authors state that IVA is not deterministic: even with greedy decoding, continuous batching, concurrent requests, reduced-precision arithmetic, and prefix caching make token-level output depend on batch composition, and evidence tools query live public resources that are updated or occasionally rate-limit or time out, so repeated runs of the same case can differ in wording, retrieved evidence, and occasionally in a secondary finding or a borderline classification. Evaluation was retrospective at two centers in one city, and the outcome measures were recovery of expert-reported variants and agreement with expert classes, not report quality or a comparison with analyst time or accuracy; precision is low because sensitivity was favored, and analysts must still review every call. The singleton cohort comes from consanguineous families enriched for homozygous recessive disease, the benchmark ranks models but is not a validated instrument, and multiple-choice performance is an imperfect proxy for clinical competence. The authors also note that adversarial or malformed inputs have not been tested and that parental allelic balance is used without read depth, so a poorly covered parent can mimic a de novo variant; such calls are retained with a recommendation for Sanger confirmation.